Add 52 public API operations, archive retrieval, and four live adapters

[kvkenyon]

This adds 52 source-backed Public Reports operations through reusable product/artifact discovery and generation, bringing typed installed-wheel retrieval to 56 unique paths of 243. It also adds four public rolling dashboard adapters, RT annual-file iteration, public archive/bundle metadata, and a seven-column Resource DME CSV decoder. Existing legacy imports, signatures, defaults, models, exceptions and DataFrame behavior are unchanged; the three legacy runtime files are byte-identical to audit base d1daad25df42d3fff41f88b907d99ef325b970e0.

Stacked on PR #3 (feat/public-rt-load, 12d60654632ce5c47543ef5f3410ec55972c8471). Do not merge as part of this task.

Exact coverage and evidence

• docs/evidence/installed-registry-index.json: 56 generated/installed unique paths; 57 current requests across 56 paths and 56 oldest-first requests across 56 paths. The new 52-path batch contributes 104 one-row requests. Every entry contains public receipt hashes/URLs/times and installed-module evidence.
• Pagination remains separate: two paths have live multiple-page evidence, and one bounded RT selection completed two pages/two rows. New paths have no live pagination or full-history completeness claim. Offline tests cover complete iteration and safeguards.
• Installed anonymous snapshots cover fuel mix, grid conditions/PRC, generation outages, and DC-tie flows. RT annual listing plus cached 2010/2026 source rows and current/oldest DME documents decode through the installed wheel.
• Public root metadata lists 116 products and 242 artifacts. Archive metadata verifies DME documents and available retired total-AS-offer/SASM history. First-page retired listings are 1,000/4,236 and 1,000/2,602 documents. A 103-entry DAM monthly bundle listing passed; one bounded oldest bundle transfer failed and no successful transfer is claimed.
• The remaining 187 observed data paths and other public document/history families remain active subsequent work, not authorized deferral. BINARY content cannot generate row models; public document retrieval remains in scope. Retired publication and available history are separate. Unknown/missing/ambiguous row schemas stay unsupported.
• Secure, Certified, participant EWS, private telemetry/bids/COP/awards, participant settlements, and customer data remain restricted. Hourly-load 2001 remains unavailable. Published delayed disclosures explicitly marked Public are a separate public category.

Captain-authorized Public API authentication occurred. No restricted request, paid access, secret copying, forecaster computation, historical bulk extraction, no-mistakes run, or merge occurred. Offline import/generation/tests need no credentials. The ignored mode-0600 .env is not committed.

Primary sources: ERCOT Public Reports OpenAPI, public product root, and the per-source ERCOT URLs/receipts checked into docs/evidence/. Generation reads pinned local projections only and never refreshes the legacy generated file from a moving URL.

Validation

Passed in the working tree:

uv run python tools/generate_client.py --check
uv run python tools/generate_public.py --check
uv run --extra files pytest -q
uv build
uv pip install --python /tmp/tinyercot-typing-proof.fbtIin/venv/bin/python --no-deps --reinstall dist/tinyercot-0.2.2-py3-none-any.whl
uv run python tools/check_typing.py --python /tmp/tinyercot-typing-proof.fbtIin/venv/bin/python
git diff --cached --check
git diff d1daad25df42d3fff41f88b907d99ef325b970e0 -- tinyercot/__init__.py tinyercot/_client.py tinyercot/_generated.py
git check-ignore .env

Full suite: 1259 passed, 1 skipped. Real installed-wheel PEP 561 proof: 510 legacy methods, 1,334 legacy row fields, 56 current endpoint/product facades, 498 current row fields, 847 filter fields, live/archive/metadata APIs, and 64 invalid uses rejected. Legacy diff empty; credential file ignored.

Concurrent follow-up work stays uncommitted. Final PR-specific validation used the staged snapshot below, after installing openpyxl==3.1.5 in both separate test environments:

git checkout-index --all --prefix=/tmp/tinyercot-pr4-check.fjDuNo/
cd /tmp/tinyercot-pr4-check.fjDuNo
/tmp/tinyercot-locked-wheel/bin/python -m pytest -q
/tmp/tinyercot-typing-proof.fbtIin/venv/bin/python -m pytest -q
/tmp/tinyercot-typing-proof.fbtIin/venv/bin/python -m ruff check tools tinyercot/catalog.py tinyercot/public tests
/tmp/tinyercot-typing-proof.fbtIin/venv/bin/python -m ruff format --check tools tinyercot/catalog.py tinyercot/public tests
/tmp/tinyercot-typing-proof.fbtIin/venv/bin/python tools/generate_client.py --check
/tmp/tinyercot-typing-proof.fbtIin/venv/bin/python tools/generate_public.py --check

Python 3.11.14 and 3.14.0: 1259 passed, 1 skipped each. Lint passed; 62 files formatted; both generators reproduced pinned output. Source/fixture inventories were checked independently. Public Python documentation was reviewed for Google-style argument, result, lifecycle and failure contracts.

docs/public-metadata-task-note.md records exact installed probe commands, source limits, and next active batches. docs/api-archive-installed-task-note.md records the bounded bundle failure. No live requests were repeated merely to assemble the receipt index. Next stacked work expands the already captured renewable/load/RUC batch and continuing public disclosure, archive, and live coverage.

CI follow-up: the first run caught stale fixed counts in standalone tests/wheel_smoke.py after generation/tests/lint/build passed. It now reconciles installed coverage with the installed registry. uv run ruff check tests/wheel_smoke.py, /tmp/tinyercot-typing-proof.fbtIin/venv/bin/python -I tests/wheel_smoke.py, and /tmp/tinyercot-locked-wheel/bin/python -I tests/wheel_smoke.py passed. Only validation/docs changed in this follow-up.